deepfake / app.py
varma123's picture
Update app.py
1f60e19 verified
#before running this please change the RUNTIME to GPU (Runtime -> Change runtime type -> set harware accelarotor as GPU)
#Mount our google drive
#Note : only needed when you have to download the processed data to the environment
#download and unzip the data from google drive Colab environment
#THis code is to check if the video is corrupted or not..
#If the video is corrupted delete the video.
import glob
import torch
import torchvision
from torchvision import transforms
from torch.utils.data import DataLoader
from torch.utils.data.dataset import Dataset
import os
import numpy as np
import cv2
import matplotlib.pyplot as plt
import face_recognition
#Check if the file is corrupted or not
def validate_video(vid_path,train_transforms):
transform = train_transforms
count = 20
video_path = vid_path
frames = []
a = int(100/count)
first_frame = np.random.randint(0,a)
temp_video = video_path.split('/')[-1]
for i,frame in enumerate(frame_extract(video_path)):
frames.append(transform(frame))
if(len(frames) == count):
break
frames = torch.stack(frames)
frames = frames[:count]
return frames
#extract a from from video
def frame_extract(path):
vidObj = cv2.VideoCapture(path)
success = 1
while success:
success, image = vidObj.read()
if success:
yield image
im_size = 112
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
train_transforms = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((im_size,im_size)),
transforms.ToTensor(),
transforms.Normalize(mean,std)])
video_fil = glob.glob('dataset final/*.mp4')
print("Total no of videos :" , len(video_fil))
print(video_fil)
count = 0;
for i in video_fil:
try:
count+=1
validate_video(i,train_transforms)
except:
print("Number of video processed: " , count ," Remaining : " , (len(video_fil) - count))
print("Corrupted video is : " , i)
continue
print((len(video_fil) - count))
#to load preprocessod video to memory
import json
import glob
import numpy as np
import cv2
import copy
import random
video_files = glob.glob('dataset final/*.mp4')
random.shuffle(video_files)
random.shuffle(video_files)
frame_count = []
for video_file in video_files:
cap = cv2.VideoCapture(video_file)
if(int(cap.get(cv2.CAP_PROP_FRAME_COUNT))<100):
video_files.remove(video_file)
continue
frame_count.append(int(cap.get(cv2.CAP_PROP_FRAME_COUNT)))
print("frames are " , frame_count)
print("Total no of video: " , len(frame_count))
print('Average frame per video:',np.mean(frame_count))
# load the video name and labels from csv
import torch
import torchvision
from torchvision import transforms
from torch.utils.data import DataLoader
from torch.utils.data.dataset import Dataset
import os
import numpy as np
import cv2
import matplotlib.pyplot as plt
import face_recognition
class video_dataset(Dataset):
def __init__(self,video_names,labels,sequence_length = 60,transform = None):
self.video_names = video_names
self.labels = labels
self.transform = transform
self.count = sequence_length
def __len__(self):
return len(self.video_names)
def __getitem__(self,idx):
video_path = self.video_names[idx]
frames = []
a = int(100/self.count)
first_frame = np.random.randint(0,a)
temp_video = video_path.split('/')[-1]
#print(temp_video)
label = self.labels.iloc[(labels.loc[labels["filename"] == temp_video].index.values[0]),1]
if(label == 'fake'):
label = 0
if(label == 'real'):
label = 1
for i,frame in enumerate(self.frame_extract(video_path)):
frames.append(self.transform(frame))
if(len(frames) == self.count):
break
frames = torch.stack(frames)
frames = frames[:self.count]
#print("length:" , len(frames), "label",label)
return frames,label
def frame_extract(self,path):
vidObj = cv2.VideoCapture(path)
success = 1
while success:
success, image = vidObj.read()
if success:
yield image
#plot the image
def im_plot(tensor):
image = tensor.cpu().numpy().transpose(1,2,0)
b,g,r = cv2.split(image)
image = cv2.merge((r,g,b))
image = image*[0.22803, 0.22145, 0.216989] + [0.43216, 0.394666, 0.37645]
image = image*255.0
plt.imshow(image.astype(int))
plt.show()
#count the number of fake and real videos
def number_of_real_and_fake_videos(data_list):
header_list = ["filename","label"]
lab = pd.read_csv('labels.csv',names=header_list)
fake = 0
real = 0
for i in data_list:
temp_video = i.split('/')[-1]
label = lab.iloc[(labels.loc[labels["filename"] == temp_video].index.values[0]),1]
if(label == 'fake'):
fake+=1
if(label == 'real'):
real+=1
return real,fake
# load the labels and video in data loader
import random
import pandas as pd
from sklearn.model_selection import train_test_split
header_list = ["filename","label"]
labels = pd.read_csv('labels.csv',names=header_list)
#print(labels)
train_videos = video_files[:int(0.75*len(video_files))]
valid_videos = video_files[int(0.75*len(video_files)):]
valid_label = labels[int(0.75*len(labels)):]
print("train : " , len(train_videos))
print("test : " , len(valid_videos))
# train_videos,valid_videos = train_test_split(data,test_size = 0.2)
# print(train_videos)
print("TRAIN: ", "Real:",number_of_real_and_fake_videos(train_videos)[0]," Fake:",number_of_real_and_fake_videos(train_videos)[1])
print("TEST: ", "Real:",number_of_real_and_fake_videos(valid_videos)[0]," Fake:",number_of_real_and_fake_videos(valid_videos)[1])
im_size = 112
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
train_transforms = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((im_size,im_size)),
transforms.ToTensor(),
transforms.Normalize(mean,std)])
test_transforms = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((im_size,im_size)),
transforms.ToTensor(),
transforms.Normalize(mean,std)])
train_data = video_dataset(train_videos,labels,sequence_length = 10,transform = train_transforms)
#print(train_data)
val_data = video_dataset(valid_videos,labels,sequence_length = 10,transform = train_transforms)
train_loader = DataLoader(train_data,batch_size = 4,shuffle = True,num_workers = 4)
valid_loader = DataLoader(val_data,batch_size = 4,shuffle = True,num_workers = 4)
image,label = train_data[0]
im_plot(image[0,:,:,:])
lst = [[1,2],[3,4],[4,5]]
val_labels = [val_data[i][1] for i in range(len(val_data))]
val_labels
# for item in range(0,1):
print(train_data[1][1])
val_data[1][1]
valid_label['label']
from torch import nn
import timm
class Model1(nn.Module):
def __init__(self, num_classes, latent_dim=2048, lstm_layers=1, hidden_dim=2048, bidirectional=False):
super(Model1, self).__init__()
model = timm.create_model('xception', pretrained=True)
print(model)
self.model = nn.Sequential(*list(model.children())[:-1])
self.lstm = nn.LSTM(latent_dim, hidden_dim, lstm_layers, bidirectional)
self.relu = nn.LeakyReLU()
self.dp = nn.Dropout(0.4)
self.linear1 = nn.Linear(2048, num_classes)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
def forward(self, x):
batch_size, seq_length, c, h, w = x.size()
print("Input tensor shape:", x.size())
# Reshape the input tensor to (batch_size * sequence_length, channels, height, width)
x = x.view(batch_size * seq_length, c, h, w)
fmap = self.model(x)
# Reshape the feature map to (batch_size, sequence_length, num_features)
fmap = fmap.view(batch_size, seq_length, -1)
x_lstm, _ = self.lstm(fmap, None)
return fmap, self.dp(self.linear1(torch.mean(x_lstm, dim=1)))
model1 = Model1(2).cuda()
a,b = model1(torch.from_numpy(np.empty((1,20,3,112,112))).type(torch.cuda.FloatTensor))
import torch
from torch.autograd import Variable
import time
import os
import sys
import os
def train_epoch(epoch, num_epochs, data_loader, model, criterion, optimizer):
model.train()
losses = AverageMeter()
accuracies = AverageMeter()
t = []
for i, (inputs, targets) in enumerate(data_loader):
if torch.cuda.is_available():
targets = targets.type(torch.cuda.LongTensor)
inputs = inputs.cuda()
_,outputs = model(inputs)
loss = criterion(outputs,targets.type(torch.cuda.LongTensor))
acc = calculate_accuracy(outputs, targets.type(torch.cuda.LongTensor))
losses.update(loss.item(), inputs.size(0))
accuracies.update(acc, inputs.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
sys.stdout.write(
"\r[Epoch %d/%d] [Batch %d / %d] [Loss: %f, Acc: %.2f%%]"
% (
epoch,
num_epochs,
i,
len(data_loader),
losses.avg,
accuracies.avg))
torch.save(model.state_dict(),'/content/drive/MyDrive/checkpoint1.pt')
return losses.avg,accuracies.avg
def test(epoch,model, data_loader ,criterion):
print('Testing')
model.eval()
losses = AverageMeter()
accuracies = AverageMeter()
pred = []
true = []
count = 0
with torch.no_grad():
for i, (inputs, targets) in enumerate(data_loader):
if torch.cuda.is_available():
targets = targets.cuda().type(torch.cuda.FloatTensor)
inputs = inputs.cuda()
_,outputs = model(inputs)
loss = torch.mean(criterion(outputs, targets.type(torch.cuda.LongTensor)))
acc = calculate_accuracy(outputs,targets.type(torch.cuda.LongTensor))
_,p = torch.max(outputs,1)
true += (targets.type(torch.cuda.LongTensor)).detach().cpu().numpy().reshape(len(targets)).tolist()
pred += p.detach().cpu().numpy().reshape(len(p)).tolist()
losses.update(loss.item(), inputs.size(0))
accuracies.update(acc, inputs.size(0))
sys.stdout.write(
"\r[Batch %d / %d] [Loss: %f, Acc: %.2f%%]"
% (
i,
len(data_loader),
losses.avg,
accuracies.avg
)
)
print('\nAccuracy {}'.format(accuracies.avg))
return true,pred,losses.avg,accuracies.avg
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def calculate_accuracy(outputs, targets):
batch_size = targets.size(0)
_, pred = outputs.topk(1, 1, True)
pred = pred.t()
correct = pred.eq(targets.view(1, -1))
n_correct_elems = correct.float().sum().item()
return 100* n_correct_elems / batch_size
import seaborn as sn
#Output confusion matrix
def print_confusion_matrix(y_true, y_pred):
cm = confusion_matrix(y_true, y_pred)
print('True positive = ', cm[0][0])
print('False positive = ', cm[0][1])
print('False negative = ', cm[1][0])
print('True negative = ', cm[1][1])
print('\n')
df_cm = pd.DataFrame(cm, range(2), range(2))
sn.set(font_scale=1.4) # for label size
sn.heatmap(df_cm, annot=True, annot_kws={"size": 16}) # font size
plt.ylabel('Actual label', size = 20)
plt.xlabel('Predicted label', size = 20)
plt.xticks(np.arange(2), ['Fake', 'Real'], size = 16)
plt.yticks(np.arange(2), ['Fake', 'Real'], size = 16)
plt.ylim([2, 0])
plt.show()
calculated_acc = (cm[0][0]+cm[1][1])/(cm[0][0]+cm[0][1]+cm[1][0]+ cm[1][1])
print("Calculated Accuracy",calculated_acc*100)
def plot_loss(train_loss_avg,test_loss_avg,num_epochs):
loss_train = train_loss_avg
loss_val = test_loss_avg
print(num_epochs)
epochs = range(1,num_epochs+1)
plt.plot(epochs, loss_train, 'g', label='Training loss')
plt.plot(epochs, loss_val, 'b', label='validation loss')
plt.title('Training and Validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()
plt.show()
def plot_accuracy(train_accuracy,test_accuracy,num_epochs):
loss_train = train_accuracy
loss_val = test_accuracy
epochs = range(1,num_epochs+1)
plt.plot(epochs, loss_train, 'g', label='Training accuracy')
plt.plot(epochs, loss_val, 'b', label='validation accuracy')
plt.title('Training and Validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend()
plt.show()
from sklearn.metrics import confusion_matrix
#learning rate
lr = 1e-5#0.001
#number of epochs
num_epochs = 40
optimizer = torch.optim.Adam(model1.parameters(), lr= lr,weight_decay = 1e-5)
#class_weights = torch.from_numpy(np.asarray([1,15])).type(torch.FloatTensor).cuda()
#criterion = nn.CrossEntropyLoss(weight = class_weights).cuda()
criterion = nn.CrossEntropyLoss().cuda()
train_loss_avg =[]
train_accuracy = []
test_loss_avg = []
test_accuracy = []
for epoch in range(1,num_epochs+1):
l, acc = train_epoch(epoch,num_epochs,train_loader,model1,criterion,optimizer)
train_loss_avg.append(l)
train_accuracy.append(acc)
true,pred,tl,t_acc = test(epoch,model1,valid_loader,criterion)
test_loss_avg.append(tl)
test_accuracy.append(t_acc)
plot_loss(train_loss_avg,test_loss_avg,len(train_loss_avg))
plot_accuracy(train_accuracy,test_accuracy,len(train_accuracy))
print(confusion_matrix(true,pred))
print_confusion_matrix(true,pred)
from torch import nn
import timm
class Model2(nn.Module):
def __init__(self, num_classes, latent_dim=2048, lstm_layers=1, hidden_dim=2048, bidirectional=False):
super(Model2, self).__init__()
# Create the Inception model
model = timm.create_model('inception_v3', pretrained=True)
# Remove the classification head
model = list(model.children())[:-1]
self.model = nn.Sequential(*model)
self.lstm = nn.LSTM(latent_dim, hidden_dim, lstm_layers, bidirectional)
self.relu = nn.LeakyReLU()
self.dp = nn.Dropout(0.4)
# Linear layer for classification
self.linear1 = nn.Linear(2048, num_classes)
# Adaptive pooling layer
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
def forward(self, x):
batch_size, seq_length, c, h, w = x.size()
print("Input tensor shape:", x.size())
# Reshape the input tensor to (batch_size * sequence_length, channels, height, width)
x = x.view(batch_size * seq_length, c, h, w)
fmap = self.model(x)
# Reshape the feature map to (batch_size, sequence_length, num_features)
fmap = fmap.view(batch_size, seq_length, -1)
x_lstm, _ = self.lstm(fmap, None)
return fmap, self.dp(self.linear1(torch.mean(x_lstm, dim=1)))
model2 = Model2(2).cuda()
a,b = model2(torch.from_numpy(np.empty((1,20,3,112,112))).type(torch.cuda.FloatTensor))
from sklearn.metrics import confusion_matrix
#learning rate
lr = 1e-5#0.001
#number of epochs
num_epochs = 50
optimizer = torch.optim.Adam(model2.parameters(), lr= lr,weight_decay = 1e-5)
#class_weights = torch.from_numpy(np.asarray([1,15])).type(torch.FloatTensor).cuda()
#criterion = nn.CrossEntropyLoss(weight = class_weights).cuda()
criterion = nn.CrossEntropyLoss().cuda()
train_loss_avg =[]
train_accuracy = []
test_loss_avg = []
test_accuracy = []
for epoch in range(1,num_epochs+1):
l, acc = train_epoch(epoch,num_epochs,train_loader,model2,criterion,optimizer)
train_loss_avg.append(l)
train_accuracy.append(acc)
true,pred,tl,t_acc = test(epoch,model2,valid_loader,criterion)
test_loss_avg.append(tl)
test_accuracy.append(t_acc)
plot_loss(train_loss_avg,test_loss_avg,len(train_loss_avg))
plot_accuracy(train_accuracy,test_accuracy,len(train_accuracy))
print(confusion_matrix(true,pred))
print_confusion_matrix(true,pred)
#Model with feature visualization
from torch import nn
from torchvision import models
class Model3(nn.Module):
def __init__(self, num_classes,latent_dim= 2048, lstm_layers=1 , hidden_dim = 2048, bidirectional = False):
super(Model3, self).__init__()
model = models.resnext50_32x4d(pretrained = True)
self.model = nn.Sequential(*list(model.children())[:-2])
self.lstm = nn.LSTM(latent_dim,hidden_dim, lstm_layers, bidirectional)
self.relu = nn.LeakyReLU()
self.dp = nn.Dropout(0.4)
self.linear1 = nn.Linear(2048,num_classes)
self.avgpool = nn.AdaptiveAvgPool2d(1)
def forward(self, x):
batch_size,seq_length, c, h, w = x.shape
x = x.view(batch_size * seq_length, c, h, w)
fmap = self.model(x)
x = self.avgpool(fmap)
x = x.view(batch_size,seq_length,2048)
x_lstm,_ = self.lstm(x,None)
return fmap,self.dp(self.linear1(x_lstm[:,-1,:]))
model3 = Model3(2).cuda()
a,b = model3(torch.from_numpy(np.empty((1,20,3,112,112))).type(torch.cuda.FloatTensor))
from sklearn.metrics import confusion_matrix
#learning rate
lr = 1e-5#0.001
#number of epochs
num_epochs = 50
optimizer = torch.optim.Adam(model3.parameters(), lr= lr,weight_decay = 1e-5)
#class_weights = torch.from_numpy(np.asarray([1,15])).type(torch.FloatTensor).cuda()
#criterion = nn.CrossEntropyLoss(weight = class_weights).cuda()
criterion = nn.CrossEntropyLoss().cuda()
train_loss_avg =[]
train_accuracy = []
test_loss_avg = []
test_accuracy = []
for epoch in range(1,num_epochs+1):
l, acc = train_epoch(epoch,num_epochs,train_loader,model3,criterion,optimizer)
train_loss_avg.append(l)
train_accuracy.append(acc)
true,pred,tl,t_acc = test(epoch,model3,valid_loader,criterion)
test_loss_avg.append(tl)
test_accuracy.append(t_acc)
plot_loss(train_loss_avg,test_loss_avg,len(train_loss_avg))
plot_accuracy(train_accuracy,test_accuracy,len(train_accuracy))
print(confusion_matrix(true,pred))
print_confusion_matrix(true,pred)
models = [model1, model2]
# preds = [model.predict(valid_loader) for model in models]
true1,pred1,tl1,t_acc1 = test(epoch,model1,valid_loader,criterion)
true2,pred2,tl2,t_acc2 = test(epoch,model2,valid_loader,criterion)
true3,pred3,tl3,t_acc3 = test(epoch,model3,valid_loader,criterion)
preds=np.array([pred1,pred2])
# summed = np.sum(preds, axis=0)
# # argmax across classes
# ensemble_prediction = np.argmax(summed, axis=1)
# # combined_pred = (pred1 + pred2) / 2 # Averaging the predictions of the two models
# # Compute combined accuracy
# combined_accuracy = (t_acc1 + t_acc2) / 2
# # print(combined_pred)
# print(combined_accuracy)
# print((pred1))
# w1 = 0.5
# w2 = 1.5
# combined_accuracy = (w1 * t_acc1 + w2 * t_acc2) / (w1 + w2)
# combined_accuracy
# summed = np.sum(preds, axis=0)
# print(summed)
# # argmax across classes
# ensemble_prediction = np.argmax(summed)
# ensemble_prediction
# # print(epoch)
# # print(model1)
# print(np.array(valid_loader))
# print(criterion)
print(val_data)
# import numpy as np
# # Convert pred1 and pred2 to numpy arrays
# pred1_array = np.array(pred1)
# pred2_array = np.array(pred2)
# val_array = np.array(val_labels)
# weighted_pred = (pred2_array + pred1_array) / 2
# # Compute accuracy of the ensemble model
# ensemble_accuracy = np.mean(np.argmax(weighted_pred) == val_array)
# print("Ensemble accuracy:", ensemble_accuracy)
import numpy as np
# Define a grid of weights to search over
weight_range = np.linspace(0, 1, num=11) # Adjust the number of values and range as needed
best_accuracy = 0.0
best_weights = None
# Iterate over all combinations of weights
for w1 in weight_range:
for w2 in weight_range:
for w3 in weight_range:
# Ensure the sum of weights is 1
total_weight = w1 + w2 + w3
if total_weight == 0:
continue
w1 /= total_weight
w2 /= total_weight
w3 /= total_weight
# Combine accuracies using weighted average
weighted_accuracy = (w1 * t_acc1 +
w2 * t_acc2 +
w3 * t_acc3)
# Update best accuracy and weights if current ensemble accuracy is higher
if weighted_accuracy > best_accuracy:
best_accuracy = weighted_accuracy
best_weights = (w1, w2, w3)
print("Best ensemble accuracy:", best_accuracy)
print("Best weights:", best_weights)
# weight_range = np.linspace(0, 1, num=11)
# weight_range